The research desk

Spec your research project.

This is a scoping request, not a quote. A researcher reads your question and comes back with what could actually answer it, what each method could and could not evidence, and what it would take. Prices are on the pricing page; nothing here commits you to any of them.

Start with the question. Everything else — who it is about, what it will change, when you need it — is what decides which methods could honestly answer it.

Please do not paste anything confidential, personal data about other people, or anything under NDA. Describe the question — the detail belongs in the conversation.

How we would answer it

Four kinds of evidence, and what each one can honestly carry.

Most research is sold on what it will show. This is the other half: what each method could never show, however the report is written. A study we run will label every finding with one of these four, and print the limits beside the number rather than in a footnote.

Measured

A deterministic check run over a real artefact. Re-running it on the same artefact returns the same answer.

Can evidence

  • That a specific artefact does or does not contain a specific thing, on a stated date.
  • Counts and rates over the artefacts actually checked (n is the artefacts, not the market).
  • Change over time, where the same check ran on the same artefacts at two dates.

Cannot evidence

  • Why anything is the way it is.
  • What anybody thinks, wants, or would do.
  • Anything about artefacts that were not checked — including the rest of the market.

Inferred

A model's reading of an artefact. A capable, fallible reader's opinion — not a measurement, and not a person's view.

Can evidence

  • A defensible reading of what a document says or implies.
  • Consistent classification across many artefacts read the same way.
  • Hypotheses and framings worth testing by another method.

Cannot evidence

  • That the reading is correct.
  • What a human reader would conclude.
  • Anything about the world the model could not see — most importantly, anything the run's own constraints excluded. (This is the solar retraction, exactly.)

Observed

What a real, identifiable-to-us-but-pseudonymous human being actually did or said, having been recruited and paid to do a task.

Can evidence

  • That these specific people, at this time, did or said this.
  • PREVALENCE in the sampled population — the share who did or said it — reported with n and a confidence interval, and only within the frame that was actually sampled.
  • What a real firm actually does when a real buyer approaches it (mystery shopping).
  • Language real buyers use, unprompted, in their own words.

Cannot evidence

  • Prevalence in any population the panel did not sample — a UK-resident consumer panel is not 'UK fleet managers'.
  • Cause, from an observational design.
  • Anything at a sample size that will not carry it. At n = 3 the interval is nearly the whole range, and that is the honest output.

Simulated

Output from a modelled population. A construct standing in for a market, made of archetype specifications and a language model. NEVER a claim about what real people think.

Can evidence

  • HYPOTHESES worth testing — the space of plausible answers, generated cheaply and at breadth.
  • LANGUAGE and framings to put in front of real people — a lexicon to test, not a lexicon in use.
  • PRIORITISATION — which questions are worth spending human fieldwork money on, and which are already settled.
  • INTERNAL CONSISTENCY of a methodology — whether the model gives stable answers to the same question across runs.
  • Coverage of edge cases and objections a small human sample would probably miss.

Cannot evidence

  • PREVALENCE. Not '62% of fleet managers say', not 'most buyers', not 'the majority'. A simulated population has no prevalence in any real population, at any sample size, ever. Running 10,000 personas does not make it a survey; it makes it a large simulation.
  • What any real person thinks, wants, would buy, or would pay.
  • Market size, share, demand, or willingness to pay.
  • That the archetypes resemble the real population — that is a separate, measured calibration against a human panel (src/lib/validation/calibration.ts), and its result is 'observed', not this.
  • Anything at all, on its own, in a published claim about a real market. Simulated output is an input to research, not a finding about the world.

The distinction that matters most is the last two. A simulated population is a construct we build to generate hypotheses cheaply and decide where to spend the fieldwork budget. It is never a claim about what real people think, and no share, percentage or majority can be read out of it — only a paid human panel can carry that, and only within the frame it actually sampled.

Published work produced this way is in the research library. Desk prices are on the pricing page.